Larrabee, Intel's once-vaunted, next-generation graphics card died years ago, but the CPU technology behind the would-be graphics card has lived on. Intel discussed the future of MIC/Knight's Corner today. After Larrabee was officially canceled, Intel repurposed the design and seeded development kits to appropriate market segments. MIC cards won't start shipping until the 22nm Knight's Corner chip is launched, but even the Knight's Ferry prototypes offer tantalizing hints at what future performance might be resemble.
Like Larrabee, Knight's Corner (and future MIC products in general) utilize a CPU based on Intel's original Pentium architecture (P54C). Modifications include complete cache coherency, x86-64 compatibility, and 512-bit vector support capable of performing 16 single-precision floating point operations simultaneously.
One advantage Intel is talking up in a big way is the fact that existing x86 applications will be relatively easy to port to MIC processors. The exact amount of necessary modification will depend on what the program does and how easy it is to parallelize the workload across the 50 chips Knight's Corner will offer. The company claims that "With greater reuse of parallel processor code, software companies and IT departments benefit from creating and maintaining a single code base binary and not having to retrain developers on proprietary programming models associated with accelerators."
The advantages of Knight's Crossing, from Intel's point of view.
That's true to a point but it's not universally applicable. Knight's Corner is almost certainly 9-12 months away, if not longer. That's more than enough time for NV and AMD to make their own move to 28nm and further evolve their own GPGPU architectures. This is more of a potential issue for Nvidia than AMD--Team Green doesn't currently have a Tesla-equivalent GPGPU product line.
x86 compatibility may give Intel an edge, but it's far from the only important factor. Knight's Corner will utilize Intel's tri-gate technology, but will rely, at least in part, on proprietary Intel development tools and the OpenMP software standard. According to SGI's CTO, Eng Lim Goh, Knight's Corner is intriguingly easy to program. Performance, however, varies. Speaking to the EETimes, Goh noted:
[P]erformance ranges dramatically based on applications from orders of magnitude improvements to incremental improvements" using the current chips [Knight's Ferry] that support only single-precision floating point operations.The one good sign (from NV's perspective) is that while Intel mentioned 'exaflop computing' multiple times in its presentation, the company refrained from publishing any benchmark results between Knight's Ferry/Knight's Corner products and any of NV's Tesla GPUs. That could mean that Intel isn't completely confident in how the two solutions stack up against each other.
You end up with two kinds of customers, one highly satisfied with [AMD and Nvidia graphics] accelerators because despite the tedious porting process, their results are very good. Others feel their time spent on porting [their apps to AMD and Nvidia chips] doesn’t justify the performance and there is a huge part of this second group for whom MIC is useful—and ultimately some of the first group may want MIC, too.
Alternatively, Santa Clara may not want to tip its hand too early. Larrabee may be dead and gone, but its legacy could still mean trouble for Nvidia.